alpaca-lora

Use lora to reproduce alpaca.

Installation

Main dependencies are openrlhf and lm-eval

Install openrlhf

Refer to openrlhf.

Install lm-eval

Refer to lm-eval-harness.

Data Generation

The data generation methodology draws inspiration from Alpaca, though with several simplifications to streamline the process. By leveraging the DeepSeek Chat API for generation, we were able to complete the asynchronous requests in approximately 17.5 hours, with the total API costs remaining under $35 USD (approximately 250 CNY).

To generate data, run the following commands:

python generate_instructions_async.py

LoRA Fine-tuning

OpenRLHF is used for easy lora fine-tuning.

./train_llama3_8b_sft_lora.sh 

Evaluation

lm-evaluation-harness is used for easy evaluation.

./eval.sh